Multi-Modal Neural Plasma Estimation on MAST-U through Synthetic Data

We develop a neural network for the estimation of spatial distributions of neutrals density, electron density and electron temperature in the MAST-U Tokamak. As neutrals are difficult to estimate experimentally and expensive to simulate, we train our model on synthetic multi-perspective camera data and sparse measurements. Prior work on training a neural model on synthetic data was limited to monocular cameras and synthetic data where its performance was inhibited by the synthetic-to-real (syn2real) domain gap, as demonstrated on idealised SOLPS simulations [1]. We overcome these limitations in our model by including multiple camera perspectives, and consider three contributions to bridge the syn2real gap. First, we use a Domain-Adversarial Neural Network [2] architecture to learn features agnostic to the real and synthetic domains. Second, we include sensor noise and calibration errors in our synthetic camera data. Finally, we use a parametric Physically Informed Neural Network [3] for physically conditioned synthetic magnetic flux maps. We show that our model is robust to imaging errors and compare the predicted emission to conventional tomographic reconstructions.

References:

1. Öztürk E et al. 51st EPS Plasma 2025

2. Ganin Y et al. JMLR 17, 1–35 2016

3. Jang B et al. POP 31, 10.1063/5.0188634 2024